A forced-choice ballot study finds sharply different candidate selections from GPT and Grok
The API experiment used search and an instruction to pick one candidate, making its testing conditions central to interpreting the results.
A Convovo evaluation asked GPT-6.1 Sol and Grok 4.7 to select one candidate in 2026 United States House, Senate and governor races through their APIs with search enabled. TechCrunch's October 24 reporting says GPT selected Democrats in 97% of answers and Grok selected Republicans in 92%, across nearly 500 races.
The prompt explicitly demanded a single recommendation and discouraged hedging. Those conditions matter: the study examines responses to a forced-choice instruction in a particular API setup. It should not be presented as the percentage of ordinary consumer-chat conversations that produce a party recommendation.
Convovo publishes examples and repeated answers for individual races, including refusals and split outcomes. TechCrunch also reports that the consumer applications initially declined to make recommendations. OpenAI told the publication that the API experiment did not represent the standard ChatGPT experience.
Search behavior offers a further point of investigation. In a later thirty-race sample without internet search, TechCrunch reports that Grok mostly refused to choose, while GPT retained a Democratic preference. A change in retrieval can therefore change the observed behavior, making the surrounding system part of the evaluation.
The practical issue is how a model selects and frames information before it gives an answer. Sources, search terms, instructions and product safeguards can all affect the result. Publishing the test conditions and individual outputs lets readers examine those influences. The study provides evidence about its tested configurations, rather than an instruction about which candidate a person should support.
Sources
- [AI Ballot2026: GPT picks Democrats. Grok picks Republicans.](https://convovo.ai/ballot/) — Convovo - [TechCrunch reporting](https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/) — TechCrunch